We killed three AI projects and saved $400,000

A company killed three AI projects and avoided an estimated $400,000 in cost.

The surprising part was not the size of the projects. It was what we found inside the work.

Sixty percent of the tasks that looked manual and repetitive should never have been automated.

The company came to me with three builds already running. The thinking was normal enough: find repetitive work, hand it to a machine, and watch the cost drop.

So we spent two days going through the jobs one by one. We watched what people actually did, not what the process document said they did.

Most of the work was not repetitive. It only looked repetitive from a distance.

Someone approving an invoice may appear to be copying a number and clicking a button. In reality, that person may be deciding whether the supplier is trustworthy, whether the amount makes sense this month, whether the purchase was authorized, and whether something feels unusual.

Automating that final decision does not automatically create a saving. It can create a fast, confident mistake.

The company stopped three projects, kept one smaller workflow, and avoided eighteen months of drift.

The figures in this article come from an anonymized client engagement. They describe this case, not an industry average.

Repetitive movement is not repetitive judgment

Most automation plans begin with a list of tasks people repeat. That is a useful start, but repetition alone is a weak filter.

Two employees may perform the same task every day while using judgment each time. The clicks repeat. The reason behind the clicks changes.

An invoice process is a good example. It contains several different kinds of work:

Step

Type of work

Better treatment

Read supplier name and amount

Data extraction

Automate

Match the invoice to a purchase order

Rule-based check

Automate

Flag a duplicate invoice

Pattern and rule check

Automate, then review

Decide whether an unusual supplier should be trusted

Contextual judgment

Keep human

Approve a high-value exception

Accountable decision

Keep human

Record the approved invoice

System update

Automate after approval

The mistake is treating the whole process as one task.

The better approach is to separate the movement of information from the judgment that carries responsibility.

The four outcomes of a proper automation audit

Every task should end in one of four groups.

Eliminate

The task creates no useful outcome, exists because of an old process, or produces a report nobody uses.

Do not automate waste. Remove it.

Simplify

The task is necessary, but the current process has extra approvals, duplicate data entry, unclear forms, or too many handoffs.

Fix the process before adding technology.

Assist

The task needs human judgment, but AI or automation can prepare the information. The system may extract data, summarize history, identify exceptions, or draft a recommendation. A person still decides.

Automate

The task follows stable rules, uses reliable data, has manageable exceptions, and creates an outcome that can be checked or reversed.

Only this final group should become a fully automated workflow.

The two-day audit we used

You can run a smaller version of this audit before approving your next AI project.

Day 1: Watch the real work

Step 1: Choose one business outcome

Do not begin with "automate finance" or "build an HR agent."

Choose one result, such as:

• Process an incoming supplier invoice
• Qualify a new sales lead
• Prepare a weekly operations report
• Route a customer support request
• Schedule an approved booking

Write where the process starts and where it ends. Anything outside that boundary belongs to another audit.

Step 2: Observe several real examples

Watch different people complete the job. Include an easy case, a normal case, and an exception.

Ask them to say what they are checking and why. Do not ask only which buttons they press.

Capture:

• The information they receive
• The systems they open
• The rules they follow
• The decisions they make
• The people they contact
• The exceptions that slow them down
• The final record or action they create

The undocumented conversations often matter more than the documented steps.

Step 3: Break the process into individual actions

Write one action per row. Avoid broad labels such as "review invoice." Break that work into reading, checking, comparing, deciding, approving, and recording.

For each action, note whether it is:

• Data movement
• A fixed calculation
• A stable rule
• Pattern recognition
• Contextual judgment
• Approval or accountability

This is where apparently repetitive work begins to separate into different risk levels.

Step 4: Find the hidden context

Ask the person doing the job:

1. When do you ignore the normal rule?
2. What makes you stop and ask someone?
3. What information do you know that is not on the screen?
4. What happened the last time this went wrong?
5. Who is responsible for the outcome?

If the answer depends on history, relationships, timing, trust, or an unusual business condition, the step contains judgment.

That does not mean technology cannot help. It means the final decision needs a person or a deterministic policy outside the model.

Day 2: Decide what deserves a build

Step 5: Score every action

Use a simple one-to-five score for each factor:

Factor

Low score

High score

Volume

Rare

Happens many times

Rule stability

Changes often

Clear and stable

Data quality

Missing or inconsistent

Structured and reliable

Exception rate

Many unusual cases

Few unusual cases

Reviewability

Hard to verify

Quick to verify

Reversibility

Damage is difficult to undo

Easy to correct

Error impact

Financial, legal, or customer harm

Small operational effect

Ownership

Nobody owns it

Named process owner

A high-volume task with stable rules, clean data, few exceptions, and reversible outcomes is a strong automation candidate.

A high-impact decision with hidden context is not, even if someone performs it every day.

Step 6: Calculate the whole cost

Do not compare the cost of software with employee time alone.

Estimate:

Build cost
+ Integration and data-cleaning cost
+ Testing and security cost
+ Expected model and platform usage
+ Monitoring and support
+ Maintenance when systems or rules change
+ Expected cost of errors
= Total operating cost

Then compare that figure with the cost of leaving the work manual, simplifying it, or using AI only to assist a person.

The cheapest option is sometimes a better form, one removed approval, or a clearer policy.

Step 7: Run the consequence test

Ask what happens when the workflow is wrong.

If a mistake creates a draft that a person reviews, the risk may be manageable. If it sends money, denies access, deletes data, changes a contract, or communicates a binding decision, the risk is much higher.

Match the control to the consequence:

Consequence

Minimum control

Internal draft is wrong

Human review

Record is created twice

Idempotency and duplicate check

Customer receives incorrect message

Approval or strict policy source

Payment or refund changes

Deterministic rules and authorized approval

Access, data, or infrastructure changes

Restricted permissions and a hard approval gate

The model should never be the only control protecting a high-impact action.

Step 8: Choose one outcome and one owner

Mark every action as eliminate, simplify, assist, or automate.

Then name the person who owns the result. A workflow without an owner slowly becomes an invisible business risk.

The owner should approve changes, review failures, track performance, and decide when the workflow must stop.

What we kept

We did not stop all automation. We kept the part that had stable inputs, clear rules, and a result people could review.

That smaller design was easier to test and cheaper to maintain. It also protected the judgment that made the process work in the first place.

This is an important distinction. Killing a weak AI project is not a failure of innovation. It is capital discipline.

The goal is not to maximize the number of AI projects. The goal is to improve the business.

Where n8n fits

Once a process has passed the audit, n8n can connect the approved steps, data sources, rule checks, alerts, and human approvals in one visible workflow.

For an invoice process, a controlled version could:

1. Receive the invoice.
2. Extract the supplier, amount, date, and purchase order number.
3. Validate the required fields.
4. Check for duplicates.
5. Match the invoice against the purchase order.
6. Flag unusual values for review.
7. Ask an authorized person to approve the payment.
8. Update the finance system only after approval.
9. Log the result and alert someone if the workflow fails.

n8n should orchestrate the approved process. It should not decide which business risks are acceptable.

Start with a manual trigger and test data. Add production credentials only after the permissions, error path, and approval gate have been tested.

The automation readiness checklist

Before funding an AI or automation project, confirm:

• The business outcome is clearly defined.
• The current process has been observed, not assumed.
• Each action has been separated from the larger job.
• Unnecessary work has been removed first.
• Rules and judgment have been identified separately.
• Hidden context and exceptions are documented.
• The data is available and reliable.
• The cost of errors is understood.
• High-impact decisions remain with an accountable person.
• The expected saving includes build, usage, support, and maintenance.
• The first version has a narrow scope.
• The workflow can be tested beside the current process.
• Failures create alerts instead of disappearing.
• The result can be corrected or reversed.
• One person owns the workflow after launch.

If several answers are missing, pause the build. Two days of diagnosis can prevent eighteen months of drift.

The rule I use

Do not automate a job because it looks repetitive.

Find the parts that move information. Find the parts that apply stable rules. Automate those first.

Keep judgment, trust, and accountability with the people who understand the consequence.

A slower human decision is often cheaper than a fast, confident mistake.

Want to find the right automation opportunities?

If you want to identify what should and should not be automated in your business, contact me through my website or send me a direct message on LinkedIn. I will help you map the work before you spend money building it.

Tool mentioned

• Build controlled workflows with n8n